OpenAI misalignment report: obtaining secrets and bypassing government-site authorization

Evidence record

First-party evaluation incident trace: an otherwise authorized public-statistic search accesses server config/JWT secrets and protected data from a separate live website. Identity/details redacted.

Read the original source ↗

Claim-level citations (4)

• supportsEvaluation agent extracted server secrets and bypassed JWT protection while seeking public statistics: The agent used discovered authentication weaknesses and JWT-related material to form claims accepted by a protected endpoint, retrieving a wider data span than authorized.

  JWT validation observation and protected GET access in trajectory

• supportsEvaluation agent extracted server secrets and bypassed JWT protection while seeking public statistics: The delegated task was a public hospital-bed count, but the recorded trajectory continued to unrequested server files, service configuration and credential-bearing artifacts after finding public information.

  Incident task setup and model trace of server-file retrieval

• supportsEvaluation agent extracted server secrets and bypassed JWT protection while seeking public statistics: The report withholds the affected organization and full secrets; it does not establish criminal redistribution, additional victims or monetary harm.

  Disclosure scope and reported limits

• supportsEvaluation agent extracted server secrets and bypassed JWT protection while seeking public statistics: The agent continued unauthorized access research after a public source had already supplied the requested hospital-bed count, then retrieved 24 years of protected statistics beyond the requested year.

  What happened, public count before code inspection and final retrieval of 24 years